AI-Based Automated Detection and Classification of Osteoporosis in Females from Dental Panoramic Radiographs.
Authors
Affiliations (3)
Affiliations (3)
- Center for Artificial Intelligence in Medicine, Imaging & Forensics and Department of Physics, School of Engineering and Sciences, Sharda University, Knowledge Park 3, Greater Noida, Uttar Pradesh, 201310, India.
- Center for Artificial Intelligence in Medicine, Imaging & Forensics and Department of Oral Medicine & Radiology, School of Dental Sciences, Sharda University, Knowledge Park 3, Greater Noida, Uttar Pradesh, 201310, India.
- Center for Artificial Intelligence in Medicine, Imaging & Forensics and Department of Physics, School of Engineering and Sciences, Sharda University, Knowledge Park 3, Greater Noida, Uttar Pradesh, 201310, India. [email protected].
Abstract
Osteoporosis is a medical condition characterized by a decrease in bone mineral density and quality. In lower income societies, relatively more expensive osteoporosis-specific tests such as dual energy X-ray absorptiometry (DEXA) are rarely performed, but dental images such as dental panoramic radiographs (DPRs) are less expensive and more common. In this study, a U-Net inspired segmentation model is developed to automatically segment a part of the lower mandible from DPRs. The segmented regions of interest (ROIs) are used to train convolutional neural networks (CNN) as feature extractors. Three classifiers, namely the DenseNet classifier, support vector machine classifier (SVC), and random forest classifier (RFC), are used on top of the CNN feature extractor. Models are trained on nearly 18,000 images generated after augmentation of the original datasets of 919 DPRs. Testing is done on 10% of the augmented images separated before training. The CNN-DenseNet model's accuracies on the unseen test dataset are found to be 97% for binary classification (osteoporosis and no osteoporosis) and 96% for three class classification (no osteoporosis, mild osteoporosis, and severe osteoporosis). The testing accuracies of the CNN-SVC model and CNN-RFC model are 94% each for binary classification and 91% and 94% for three class classification, respectively.